Aikido evaluated cyber AI with 11.7B tokens and proved specialized vulnerability models find more code bugs at 50% lower cost than general closed models. General models have high false positives. Specialized vulnerability scanners secure code faster for less compute.